Deep Learning ANN Classification Model

Hi,

I am trying to develop a pattern recognition classification ANN model for 10 different classes using 11 inputs. The model runs but the performance is poor (less than 20%). I want to try to a deep learning technique. However, all the examples are for image-based classification. My problem is numerical-based (no image). Is there a way to do a deep learning pattern-based classification model in MATLAB? Is there an example on how to do that?

Answer 1

[Cris LaPierre]

Absolutely. Here is a page showing multiple examples, none of which are images: https://www.mathworks.com/help/deeplearning/gs/pattern-recognition-with-a-shallow-neural-network.html

These are all shallow networks. You can turn a shallow network into a deep network by adding more layers.

I would suggest using the Neural Network Pattern Recognition App to create a network, and then export the code. You can then manually expand that. Here's an example I built up based on the Iris data set example in the app (4 inputs, 3 outputs)

% Solve a Pattern Recognition Problem with a Neural Network
% Script generated by Neural Pattern Recognition app
% Created 05-Dec-2023 10:55:42
%
% This script assumes these variables are defined:
%
%   irisInputs - input data.
%   irisTargets - target data.

x = irisInputs;
t = irisTargets;

% Choose a Training Function
% For a list of all training functions type: help nntrain
% 'trainlm' is usually fastest.
% 'trainbr' takes longer but may be better for challenging problems.
% 'trainscg' uses less memory. Suitable in low memory situations.
trainFcn = 'trainscg';  % Scaled conjugate gradient backpropagation.

This is the section of code that creates the layers. It is a shallow network because there is only 1 hidden layer

% Create a Pattern Recognition Network
hiddenLayerSize = 10;
net = patternnet(hiddenLayerSize, trainFcn);

You can turn this into a deep learning network by adding more hidden layers. For example, this code would create a 3-layer network.

% Three hidden layer NN

hiddenLayerSize1 = 10;

hiddenLayerSize2 = 20;

hiddenLayerSize3 = 15;

net = patternnet([hiddenLayerSize1 hiddenLayerSize2 hiddenLayerSize3], trainFcn);

% Setup Division of Data for Training, Validation, Testing
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

% Train the Network
[net,tr] = train(net,x,t);

% Test the Network
y = net(x);
e = gsubtract(t,y);
performance = perform(net,t,y)
tind = vec2ind(t);
yind = vec2ind(y);
percentErrors = sum(tind ~= yind)/numel(tind);

% View the Network
view(net)

% Plots
% Uncomment these lines to enable various plots.
%figure, plotperform(tr)
%figure, plottrainstate(tr)
%figure, ploterrhist(e)
%figure, plotconfusion(t,y)
%figure, plotroc(t,y)

[Cris LaPierre]

You'll need to load the Iris data set to run this code.

% Load the Iris dataset
[IrisInput,IrisOutput] = iris_dataset;

IrisInput contains 150 observations with four input features corresponding to four measurements:

• Sepal length in cm
• Sepal width in cm
• Petal length in cm
• Petal width in cm

IrisOutput is a 3-by-150 matrix that indicates the classification of each observation by placing a 1 in the row corresponding to the correct species: first fow for setosa, second row for versicolor, or third for virginica.

For more information on this dataset, run the following help command

help iris_dataset

[Mostafa]

Thanks for your answer. It was of great help. Unfortunately, while the accuracy did improve a little bit, it is still low (27% from the confusion matrix). Any ohter suggestions on how to improve the accuracy? I had some inputs that were 0 and 1 (yes or no) and I tried to remove them thinking that they may be some prediction challenges. It did help but nothing that significant.

[Mostafa]

I actually found that MATLAB does not define my output classes correctly. I have 34,610 observations divided into 10 output classes. However, when I look at the confusion matrix, some of the output classes have the term NaN with zero observation. To define the output classes, I used binary codes. For example, the first output class is:

0 0 0 0 0 0 0 0 0 1

The second is:

0 0 0 0 0 0 0 0 1 0

Any ideas what may be causing this?

[Mostafa]

As an update, I reran the file with only 219 observations and I am facing the same problems. Attached is the confusion matrix.

[Cris LaPierre]

Consider attaching your data, and share your code. Without that, we can only talk generally.

[Mostafa]

Here is the code and small data set. I imported the data and transposed them.

[Cris LaPierre]

Not knowing anything about your data, I think you may need to look into feature engineering. Three of your inputs are highly correlated, meaning they aren't adding anything new to the model. However, even after eliminating them, I get similar results. To me, this means there is not enough difference in your inputs to generate an accurate model.

It = readmatrix("InputsMethod1.xlsx");
Ot = readmatrix("Outputs1.xlsx");
xnames = "input" + (1:4);

x = It';
t = Ot';

% turn t into vector of 'class labels'
f = max((0.1:0.1:1)' .* t);

% Normalize data
[x,ps]=mapminmax(x,0,1); 

figure
plot([x;f])

% Inputs 1, 2 and 3 are highly correlated
figure
gplotmatrix(It,[],f,[],[],[],[],[],xnames)

% Detemine which features have the highest predictive power
[idx,scores] = fscmrmr(It,f')

idx = 1×4

1 4 2 3

scores = 1×4

0.0397 0.0034 0.0031 0.0205

bar(scores(idx))
xlabel('Predictor rank')
ylabel('Predictor importance score')
xticklabels(xnames(idx));

Here are the results I get with a shallow network for the 2 top features.